Why Factors Facilitating Collusion May Not Predict Cartel Occurrence - Experimental Evidence - Miguel A. Fonseca, Yan Li, Hans-Theo Normann - DICE

Page created by Arnold Henderson
 
CONTINUE READING
No 289

         Why Factors Facilitating
         Collusion May Not Predict
         Cartel Occurrence –
         Experimental Evidence

         Miguel A. Fonseca,
         Yan Li,
         Hans-Theo Normann
         May 2018
IMPRINT

DICE DISCUSSION PAPER

Published by

düsseldorf university press (dup) on behalf of
Heinrich‐Heine‐Universität Düsseldorf, Faculty of Economics,
Düsseldorf Institute for Competition Economics (DICE), Universitätsstraße 1,
40225 Düsseldorf, Germany
www.dice.hhu.de

Editor:

Prof. Dr. Hans‐Theo Normann
Düsseldorf Institute for Competition Economics (DICE)
Phone: +49(0) 211‐81‐15125, e‐mail: normann@dice.hhu.de

DICE DISCUSSION PAPER

All rights reserved. Düsseldorf, Germany, 2018

ISSN 2190‐9938 (online) – ISBN 978‐3‐86304‐288‐2

The working papers published in the Series constitute work in progress circulated to
stimulate discussion and critical comments. Views expressed represent exclusively the
authors’ own opinions and do not necessarily reflect those of the editor.
Why Factors Facilitating Collusion May Not Predict
          Cartel Occurrence — Experimental Evidence∗

              Miguel A. Fonseca†, Yan Li‡ and Hans-Theo Normann§

                                            May 2018

                                              Abstract

      Factors facilitating collusion may not successfully predict cartel occurrence: when a
      factor predicts that collusion (explicit and tacit) becomes easier, firms might be less
      inclined to set up a cartel simply because tacit coordination already tends to go in hand
      with supra-competitive profits. We illustrate this issue with laboratory data. We run
      n-firm Cournot experiments with written cheap-talk communication between players
      and we compare them to treatments without the possibility to talk. We conduct this
      comparison for two, four and six firms. We find that two firms indeed find it easier
      to collude tacitly but that the number of firms does not significantly affect outcomes
      with communication. As a result, the payoff gain from communication increases with
      the number of firms, at a decreasing rate.

JEL classification: L42, C90, C70
Keywords: cartels, collusion, communication, experiments, repeated games.

  ∗
     Comments of two anonymous referees greatly improved the paper. We are also grateful to seminar
participants at the Winter Seminar 2017 (Montafon) and EARIE 2017 (Maastricht) for useful suggestions.
   †
     University of Exeter Business School, Streatham Court, Rennes Drive, Exeter EX4 4PU, UK; tel: +44
1392 262584, email: m.a.fonseca@exeter.ac.uk
   ‡
     University of Liverpool Management School, Chatham Building, Chatham Street, Liverpool L69 7ZH,
UK; Tel: +44 151 7950689; email: yan.li3@liverpool.ac.uk.
   §
     Corresponding Author. Duesseldorf Institute for Competition Economics, Heinrich-Heine-Universitaet
Duesseldorf Universitaetsstr. 1, 40225 Duesseldorf, Germany; tel: +49 211 8115297, email: nor-
mann@dice.hhu.de.
2

1       Introduction
Lists of factors facilitating collusion play a popular role in the industrial organization liter-
ature and in antitrust policy.1 Typical items on those lists include the fewness of firms (or
industry concentration),2 product homogeneity, firm symmetry or regular orders. For any
of these factors (and others), the notion is that, other things being equal, collusion is more
likely. The impact of such factors is very intuitive and can be rigorously derived with simple
repeated-game analysis.
         Despite their popularity, the power of factors facilitating collusion in predicting cartel
occurrence is limited. Empirical research studying whether the factors correlate with the
frequency of detected cartels (Posner 1970, Hay and Kelly 1974, Grout and Sonderegger
2005, Levenstein and Suslow 2006) do not report clear-cut results. This is particularly the
case for the alleged correlation between cartel frequency and concentration or the number
of firms.3 It appears that facilitating factors are by no means reliable structural indicators
of cartel occurrence, although as Stigler (1970) suggests, it may be that samples of detected
cartels are biased in one way or another.
         Why do facilitating factors not reliably predict cartel frequency? Any facilitating
factor may apply to both explicit (cartel-like) agreements and implicit (tacit) coordination.
The repeated-game incentive constraint—say, of a trigger strategy equilibrium—is a neces-
sary condition for cooperation to emerge as a subgame perfect equilibrium in both settings.
If, for example, collusion is easier with fewer firms, this will be true for legal tacit coordina-
tion and for illegal explicit cartels. But why should firms be more inclined to engage in illegal
price-fixing when they find it easier to cooperate tacitly? Instead, there may well be fewer
cartels in concentrated industries, not more; or this relationship may be non-monotonic.
         Put differently, the reason why facilitating factors are not good predictors of cartel
activity is that the decision to set up a cartel should be driven by the additional profit the
cartel leads to, taking into account fines and the foregone profit when firms do not talk.
How big the additional profits from cartel-like communication are over and above the profit
    1
      See, for example, Scherer (1980, ch. 7 and 8), Tirole (1989, ch. 6), Martin (2001, ch. 10), Motta (2004,
ch. 4.2) and Belleflamme and Peitz (2015, ch. 14). See also the treatment by Ivaldi et al. (2003).
    2
      Concentration and number of firms will be correlated but they may not always be addressed by the
same facilitating factor: more concentrated industries may be less symmetric and therefore, all else equal,
less prone to collusion.
    3
      To illustrate this puzzling finding, we quote here from various studies. Posner (1970) concludes that
“[a] large proportion [of the cartels were] in industries not normally regarded as highly concentrated”, Hay
and Kelly (1974) find that “in many cases larger groups conspire.” In a report for the British OFT, Grout
and Sonderegger (2005) state about several facilitating factors that “[i]ndeed, there is some element of
disconnection between the predictions and the variables that are relevant here.” And Levenstein and Suslow
(2006) find “no simple relationship between industry concentration and the likelihood of collusion.”
3

obtained from tacit agreements is an entirely different question. Ex ante, it is not clear at
all that facilitating factors are a good predictor of this extra margin.
        Figure 1 illustrates the issue. The literature gives the impression that, the fewer the
firms, the more explicit collusion is to be expected (downward sloping line). This conclusion
is consistent with the idea that fewer firms will find it easier to collude. However, it ignores
that explicitly colluding is costly. The costs of a cartel include the opportunity cost from
coordinating tacitly, organizational costs and cartel fines. We will argue that while more firms
may benefit more strongly from explicitly talking, the gain from talking might eventually
decline such that a medium number of firms benefits the most from colluding (concave curve).

            Likelihood of
         explicit collusion

                                                                     This paper

                                                                     Literature gives
                                                                     impression

                                  Most            Most                   Number of firms
                                collusion       collusion

                Figure 1: The number of firms and the likelihood of collusion.

        In this paper, we demonstrate the force of this argument with data from laboratory
experiments. We study one facilitating factor, namely the number of firms, and we demon-
strate with these data that such facilitating factors are not suitable for cartel detection. We
run n-firm repeated Cournot oligopoly experiments with and without communication with
the goal of measuring the gain from communication as a function of n. Specifically, markets
with two, four, and six firms which either cannot communicate at all or communicate via a
messenger-type tool. A between-subjects comparison of the profits earned for each n then
quantifies the gain from communication.
        Evidence for communication makes a fundamental difference in cartel cases, but eco-
nomic theory is currently not well equipped to justify this distinction or explain exactly how
communication facilitates cartel coordination. Regarding the notion that communication is
the defining element of a cartel, Whinston (2008) notes that “[i]t is in some sense paradoxical
4

that the least contested area of antitrust is perhaps the one in which the basis of the policy
in economic theory is weakest. ... it would be good if we understood better the economics be-
hind this”. In our experiments we explore how and to what extent communication supports
collusion by quantifying of the gain from communication in Cournot oligopoly.
        We believe that laboratory experiments are useful to examine “tacit vs. explicit col-
lusion.” In the lab we control the communication conditions rigorously. While we do not
consider lab experiments to be a substitute for field data cartel studies, it seems to us that
the polar cases of no communication and communication occur in the lab in an clean manner,
which is difficult to match with other types of data.
        We find that duopolists indeed find it easier to collude without communicating than
subjects in markets with more firms. But since the number of firms does not significantly
affect outcomes with communication, the payoff gain from communication increases with
the number of firms at a decreasing rate. That is, markets with more firms gain less from
explicit communication.

2       Literature
The literature has firmly established that communication facilitates cooperation in social
dilemmas. Deutsch (1958) finds in a prisoner’s dilemma experiment that communication
before the start of the game leads to more cooperation. One of the earliest contributions
to experimental economics, Friedman (1967), reports on a number of Cournot duopoly ex-
periments where communication is allowed, and finds that subjects often coordinate on the
joint-profit maximum. Since then, many studies of this type have confirmed this result.
Examples include Isaac, Ramey and Williams (1984), Isaac and Walker (1988), Cason and
Davis (1995) and Davis and Holt (1998). Further research has established that the exact
form of communication is important (Brosig et al. 2003), but, overall, it has been established
that talking helps.4
        Whereas there are several studies analyzing how the number of players affects the
degree of cooperation (Fouraker and Siegel 1963, Dolbear et al. 1968, Davis and Holt 1994,
Huck et al. 2004), little work with communication has been done in this area. Binger et al.
(1990) compare two and five firms in Cournot markets with and without communication.
Their results are difficult to compare to more recent studies because subjects communicated
face-to-face. Waichmann, Requate and Siang (2014) compare Cournot markets with two and
three firms using both students and managers as subjects. In addition to free communication,
    4
    Landeo and Spier (2009) demonstrate anticompetitive effects of communication in the context of exclusive
dealing. See also Boone et al. (2014).
5

they also investigate a more standardized form of chat (preformulated messages). They find
that students are affected by the type of communication whereas managers are not. Under
standardized communication, managers select lower outputs than students, but there are no
differences in subject pools under free communication. Finally, they observe more collusion in
the duopolies than in the triopolies. Harrington et al. (2016) study the effect of firm numbers
(mainly two vs. three) in markets with price-setting firms, with and without communication.
Following Holt and Davis (1990), they also study (non-binding) price announcements as
an intermediate form of communication. They find that explicit communication leads to
near-monopoly prices throughout. Announcements have only moderate effects, and only for
duopolies.
        Balliet (2010) conducts an interesting fully-fledged meta study of the effects of com-
munication in dilemma games. The effect sizes he reports are the mean differences in coop-
eration between no communication and communication. Regarding our research question,
Balliet (2010) finds that the effect of communication is stronger in larger groups. We will
return to this result in our conclusion.5
        The closest paper to our present study is Fonseca and Normann (2012), which ana-
lyzes the gain from communication for symmetric Bertrand oligopolies in lab experiments.
They conduct experiments with two, four, six, and eight firms and find an inversely u-shaped
relationship between the number of firms and the incentive to collude. Our study differs from
Fonseca and Normann (2012) in two aspects. First, we analyze strategic substitutes (homo-
geneous Cournot competition) whereas they analyze strategic complements (homogeneous
Bertrand). A recent study by Mermer, Müller and Suetens (2016) shows for two-player ex-
periments without communication that collusive outcomes differ significantly between the
two formats, which motivated us to study the environment with collusion. Second, we use
novel text-mining methods to perform an in-depth analysis of the communication between
players, thus making a contribution to the literature on pre-play communication in games.

3       Experimental design and hypotheses
We run Cournot oligopoly experiments with an inverse demand function of p = 100 − Q.
Firms have marginal costs of c = 1. We selected this set of parameters for two main reasons:
comparability to Huck et al. (2004), and the fact that they made the computation of payoffs
easy to subjects in the absence of a payoff table.6
    5
     Balliet (2010, p. 52) concedes that his study had too few larger groups to provide a thorough test and
that future research would benefit from studying larger groups. This is what our experiment does.
   6
     Designing an experiment with varying the number of players introduced tough methodological issues.
Varying the number of firms means that using a payoff table would have forced us to restrict the set of
6

       We run a 3×2 factorial design, summarized in Table 1. The first treatment variable is
the number of firms, n. We use oligopolies with n ∈ {2, 4, 6} firms.7 Our second treatment
variable is the opportunity to communicate. In the treatments without communication
(labeled No-Chat), subjects had to post quantities in each period without being able to
communicate with each other. In the communication treatments (labeled Chat), subjects
were allowed to communicate in each period via typed messages, using an instant-messenger
communication tool. Communication was unrestricted and subjects were allowed to exchange
as many messages as they liked. However, they were not allowed to identify themselves. The
time to communicate was limited to one minute in the first period and 30 seconds thereafter.

                                                          Communication
                                Number of firms           no        yes
                                                      No-Chat-2      Chat-2
                                n=2
                                                       9 (18)        9 (18)
                                                      No-Chat-4      Chat-4
                                n=4
                                                       9 (36)        9 (36)
                                                      No-Chat-6      Chat-6
                                n=6
                                                       6 (36)        6 (36)

Table 1: 3×2 factorial treatment design, treatment labels, and number of markets (and
participants) for each treatment cell.

       The experiments were implemented as a repeated game. There was a minimum num-
ber of 20 periods; after period 20, play continued for another period with a 5/6 probability.
The continuation procedure was implemented with a random computer draw. The actual
number of periods was determined ex ante and was the same in all sessions and treatments
(namely 24). Each subject participated in one repeated game only. Players were always
matched with the same partner (fixed matching).
       Table 2 summarizes the numerical predictions. In the Appendix, we provide a more
general analysis; we also do a comparative-statics analysis of n. Our first benchmark are
the static Nash equilibrium predictions (first row). If firms successfully coordinate on the
symmetric joint-profit maximum, quantities in the second row will materialize. Rows three
and four contain the profits corresponding to the static Nash equilibrium and the symmetric
joint-profit maximum.

quantities to be small. For instance, giving firms only three output levels, would result in a overly large
payoff table as there are rather many potential quantities a firm’s five competitors may produce. Furthermore,
having different treatments with different payoff table sizes could have introduced a confound by making
some treatments cognitively harder.
   7
     Our hypotheses suggest that these are indeed the relevant treatments. Moreover, in treatments with, say,
eight or more firms, even minor fluctuations around the static Nash equilibrium may end up with subjects
incurring losses.
7

                                                           n=2        n=4         n=6
                     Static Nash qi                         33.00     19.80       14.14
                     Symmetric collusive qi                 24.75     12.38        8.25

                     Static Nash Πi                      1089.00 392.04          200.02
                     Symmetric collusive Πi              1225.13 612.56          408.38

                     Minimum discount factor, δ            0.529      0.610      0.671
                     Gain from talking, ∆Π                272.25     882.09    1250.13

Table 2: Predictions. Notes: “Nash qi ” refers to the firm-level output in the one-shot
equilibrium, “Collusive qi ” refers to firm-level output in the symmetric joint-profit maximum,
δ is the minimum discount factor required in a repeated game with Nash trigger, and the
“Gain from talking” refers to the extra profit firms earn when colluding summed across all
firms.

        The conventional wisdom that fewer firms will find it easier to collude can be derived
formally in the repeated game (see the Appendix). Row five of Table 2 shows the minimum
discount factor, δ, required to sustain the symmetric joint-profit maximum as a subgame
perfect Nash equilibrium in the infinitely repeated game. This incentive constraint is a
condition that necessarily has to be met for the collusive outcome to be subgame perfect,
regardless of whether players communicate explicitly. Furthermore, such incentive conditions
are often interpreted as an indicator of how “difficult” collusion is. Thus we have theoretical
support for:

Hypothesis 1. The fewer the firms, the easier they find it to collude both (i) tacitly and
(ii) explicitly.

       Experimental evidence as well as antitrust practice suggests that firms benefit from
talking (see the introduction and the end of this section). We thus formally hypothesize that
the gain from talking is positive:

Hypothesis 2. Communication has a collusive effect.

         We now turn to the main point of the paper, the gain from communicating. Let πiChat ,
i = 1, ..., n denote the profit each firm makes when engaging in explicit communication, and
let πiN o−Chat , i = 1, ..., n, denote the profit without explicit communication.8 Then the gain
   8
     See, for example, Aubert et al. (2007) for just such a model of cartel formation. In their model, Cartel-
like communication is detected and fined with a certain probability. We will keep these factors outside our
model.
8

from communication for a market of n firms is
                                          n
                                          X
                                                 πiChat − πiN o−Chat .
                                                                    
                                   ∆Π =                                                              (1)
                                           i=1

In words, this is the amount of money the firms in an industry would put on the table in
order to be able to talk.
        Standard repeated-game theory is probably not well-equipped to predict a gain from
communicating (Harrington 2008, Whinston 2008). The incentive constraint of the repeated
game (see Appendix) merely reflects the incentives to deviate from a given collusive equilib-
rium. Whether firms coordinate on such an equilibrium with or without communication is
immaterial. Importantly, even if it turned out that πiChat and πiN o−Chat decline in the number
of firms, this does not suggest a relationship between n and ∆Π. The difference between two
monotonically declining functions can be anything, so ∆Π could be increasing, decreasing
or non-monotonic in n.
        In order to get more structure into this problem, we assume that without communi-
cation firms do not manage to sustain collusive output levels at all whereas they perfectly
collude on the monopoly output when they are allowed to talk. If so, the profits in rows three
and four of Table 2 would occur and we can calculate ∆Π (see row six, and the Appendix
for a general analysis). We formalize:

Hypothesis 3. The gain from talking, ∆Π, (i) increases monotonically in n, and (ii) it does
so at a decreasing rate.

        Note again how δ and ∆Π capture the ambiguous and apparently contradictory notion
of “facilitating collusion”. The minimum discount factor indeed suggests that fewer firms
find it easier to collude. The gain from explicit chat, however, is higher for four and six
Cournot firms than for duopolies.9 Therefore, even though fewer firms may find it easier to
collude, this does by no means imply that there will be more cartels with fewer firms.
        To what extent do we need to modify our hypotheses on ∆Π in light of previous
Cournot experiments? For Cournot markets with communication, Huck et al. (2004) found
that duopolies show some level of tacit coordination, but oligopolies with four or more
firms converge to the Nash equilibrium or are even more competitive.10 Evidence on n-firm
Cournot oligopoly with communication is less abundant but Normann et al. (2015) analyze
three-firm Cournot markets with (unstructured) communication; they report near perfect
   9
                                         P N o−Chat                       P Chat
     With perfect Bertrand competition,     πi        would be zero and      πi    would be at the profit
maximum. But then ∆Π would be constant, regardless of n.
  10
     Consistent with the experimental evidence, Li and Lyons (2012) find for telecommunication industries
that market structures with more than three firms lead to major improvements in competitiveness.
9

monopolization. Gomez-Martinez et al. (2016) report lab experiments with differentiated
Cournot competition and find that subjects in four-firm markets cooperate close to the
joint-profit maximizing level. Waichmann, Requate and Siang (2014) confirm the above
results without communication but observe less collusion with talk. While acknowledging
that we know little about Cournot markets with more than four firms, it appears the existing
experimental evidence strengthens our hypothesis that ∆Π increases in n.

4     Procedures
We provided written experimental instructions which informed subjects of all the features
of the market (the instructions are available in the Appendix). Subjects were told they were
representing one of two, four or six firms, respectively, in a market. The instructions notified
the participants of the market parameters in an informal manner. Two concrete examples
illustrated the profit calculations.
        In every period, subjects had to enter a quantity ∈ {0, 100} in a computer interface.11
Once all subjects had made their decisions, the period ended and a screen displayed the
quantity choices of all firms and the market price. The screen also displayed the individual
payoff of the current period and the accumulated payoffs up to that point but not the payoffs
of the other firms.
        Treatments were incentivized and payments were made as follows. Since losses are
possible in this game, we decided to give subjects an initial capital corresponding to four
euros. We used an experimental currency unit (“Taler”) and different exchange rates for
each market, namely 2,000 Taler for one euro for the duopolies, 1,000 Taler for the four-firm
markets and 750 Taler for the six-firm markets. The varying exchange rates are warranted
here because the pie (or the market size) is constant in this experiment whereas the number
of players is not. Payments were made at the end and in cash and consisted of the initial
capital and the sum of the payoffs attained during the course of the experiment.
        Subjects were recruited from a pool of potential participants using the online system
ORSEE (Greiner 2015). The experiments were computerized, using z-Tree (Fischbacher
2007), and were conducted at the DICElab of Heinrich-Heine University in 2013 and 2014.
A total of 180 subjects participated in 10 sessions (two duopoly sessions and four sessions each
for the n = 4 and n = 6 treatments). Sessions lasted between 45 and 65 minutes. Average
earnings were 15.39 euros and ranged from 9.26 euros (No-Com-6) to 18.86 (Com-2).
  11
     As is well known, there are additional equilibria in Cournot oligopoly when the action space is discrete
(Holt 1985). These additional equilibria are close to the prediction made in (3) (at most one unit distance)
and, moreover, may imply the same average quantities. For example, with n = 2, q1 = 34 and q2 = 32 are
mutual best replies but, the average is 33, as with continuous actions.
10

5     Results
5.1    Treatment Effects
Table 3 reports average quantities and profits conditional on the number of firms and whether
or not communication was possible. It also reports the Cournot-Nash benchmark and the
symmetric joint-profit maximum.
        Average quantities in the No-Chat condition are significantly below the Cournot-Nash
level in n = 2 (z = 2.666; p = 0.008, Wilcoxon signed-rank test (WSR); not statistically
different from Nash in n = 4 (z = −0.770; p = 0.441, WSR test) and slightly above Nash for
n = 6 (z = 1.782; p = 0.075, WSR test). These findings are consistent with results reported
in Huck et al. (2004).

                          n=2                    n=4                     n=6
                     No-Chat  Chat          No-Chat Chat            No-Chat Chat

        qi            28.86      24.84        19.22     12.01        14.94       9.01
                      (2.95)     (0.86)       (2.35)    (0.90)       (0.74)     (0.95)
        prediction    33.00      24.75        19.80     12.38        14.14       8.25

        πi           1130.66    1214.45       386.26    596.41       164.23    382.19
                     (76.29)     (9.62)      (110.82)   (24.59)      (33.85)   (36.15)
        prediction   1089.00    1225.13       392.04    612.56       200.02    408.38

        # obs           9          9            9          9            6         6

Table 3: Average quantities and profits, (std. dev.) and prediction (static Nash for No-Chat,
symmetric joint-profit maximum for Chat).

       We further observe a positive relationship between the number of firms and industry
output. A Jonckheere-Terpstra test (JT) rejected the null of the joint equality of outputs
against an ordered alternative in either direction (J ∗ = 2.089, p = 0.038). This is consistent
with Hypothesis 1 (i).

Observation 1 (i): Without communication, the fewer the firms, the easier it is to collude
tacitly.

        With communication, average quantities were very close to, and not statistically
different from, the collusion benchmark in all market structures (n = 2: z = 0.771, p = 0.441;
n = 4: z = −1.007, p = 0.314; n = 6: z = 1.572, p = 0.116, WSR, test). Along the
11

same lines, we no longer observe any statistically significant relationship between industry
output and the number of firms (JT, J ∗ = 0.000, p = 0.500, for either of the two ordered
alternatives).

Observation 1 (ii): With communication, the number of firms does not affect the level of
collusion.

        Consistent with Hypothesis 2, allowing participants to communicate leads to a sig-
nificant reduction in average quantities across all treatments (n = 2: z = 2.475, p = 0.013;
n = 4: z = 3.576, p < 0.001; n = 6: z = 2.882, p = 0.004, Wilcoxon rank-sum, WRS, test).
Consequently, firms earned higher profits in the Chat conditions than the No-Chat condi-
tions (n = 2: z = 2.209, p = 0.027; n = 4: z = 3.488, p < 0.001; n = 6: z = 2.882, p = 0.004,
WRS test).

Observation 2: The opportunity to communicate leads to lower quantities and higher
profits.
                 35
                 3025
              Mean
             20  15
                 10

                        0                            5                            10    15
                                                          Standard Deviation
                        x denotes NoChat treatments; diamond denotes Chat treatments.
                        Black: N=2; Dark Grey: N=4; Light Grey: N=6.
Figure 2: Mean and standard deviation of quantities. Observations of the No-Chat treat-
ments are denoted by ×, and  denote observations in the Chat treatments. Black symbols
refers to n = 2, dark gray to n = 4 and light gray to n = 6

        Careful observation of Table 3 shows that communication not only decreases average
firm quantities but it also reduces their dispersion. Figure 2 provides visual confirmation of
this, each (independent) group being one observation. Conditional on group size, average
12

quantities are lower under communication, and dispersion is also lower. This suggests that
communication allowed subjects to coordinate more easily on a vector of quantities, as
opposed to the case where communication was not allowed.
        Another metric of collusion is the rate at which firms were able to coordinate on the
symmetric joint-profit maximizing output, see Table 4. Referring to this output as qiC and
given the prediction is not always an integer, we define an outcome as perfectly collusive if
                                                                                   
all firms in a market post quantities within one unit of qiC , that is, qiC ≤ qi ≤ qiC ∀i.12
In the absence of communication, only duopolists managed to coordinate on the joint-profit
maximum outcome, and even then, only about a quarter of the time. (This figure is very close
to what Mermer, Müller and Suetens, 2016, observe for their strategic-substitutes duopolies).
In the four-firm and six-firm markets, firms never achieved perfect collusion. In contrast,
when communication was available, coordination on the joint profit maximizing quantity
was much more frequent. The greatest coordination benefit from communication is drawn
from four-firm markets.

                                          n=2           n=4           n=6
                           No-Chat      49 23%   0 0% 0 0%
                           Chat        137 63% 144 67% 71 49%
                           Gain         88 40% 144 67% 71 49%

                 Table 4: Absolute and relative frequency of perfect collusion

5.2     Dynamics
We now examine how quantity choices changed over the course of the experiment. We
begin by looking at whether or not individual players responded to quantities posted by
other players in the previous round, and how communication affected this. To do this, we
estimated the following Random Effects GLS model with robust standard errors, clustered
at the market level.13

  qi,t = β0 + β1 Chat + β2 Q−i,t−1 + β3 (Chat × Q−i,t−1 ) + β4 t + β5 (t × Chat) + εi,t , (2)

where qi,t is the output chosen by player i in round t of the experiment, Chat is a dummy
variable for sessions in which communication was allowed between participants, Q−i,t−1 is
  12
     Widening this interval by one unit does not change the qualitative pattern of results in terms of the
relative gains from communication.
  13
     The choice of the Random Effects estimator was driven by our use of time-invariant regressors.
13

       DV: qi,t                     n=2                     n=4                       n=6
                              (1)         (2)         (3)          (4)          (5)         (6)
       Chat                -6.28∗       -6.34∗    -12.60∗∗∗    -12.17∗∗∗      -6.97∗∗∗   -8.45∗∗∗
                           (3.37)       (3.62)      (2.88)       (2.69)        (2.41)     (2.32)
       Q−i,t−1             0.23∗∗∗      0.23∗∗∗      0.05         0.05          0.02        0.01
                           (0.08)       (0.08)      (0.03)       (0.04)        (0.03)     (0.03)
       Chat × Q−i,t−1       0.12          0.12     0.18∗∗∗      0.17∗∗∗         0.04       -0.04
                           (0.11)       (0.11)      (0.06)       (0.05)        (0.03)     (0.03)
       t                                 0.002                    0.06                   -0.06∗∗∗
                                        (0.06)                   (0.08)                   (0.02)
       t × Chat                          0.005                    -0.02                   0.10∗∗
                                        (0.07)                   (0.08)                   (0.04)
       Constant            22.29∗∗∗    22.25∗∗∗    16.35∗∗∗    15.82∗∗∗       13.45∗∗∗   14.55∗∗∗
                            (2.67)      (2.77)      (2.40)       (2.32)        (2.31)     (2.26)
       N                     828          828       1,656         1,656        1,656      1,656
       R2                    0.17         0.17       0.32          0.33        0.18       0.18
       ∗∗∗ ∗∗ ∗
         , , : respectively p < 0.01, p < 0.05, p < 0.10.
       Robust standard errors clustered at the market level in parentheses.

                      Table 5: Random effects GLS estimates of quantity

the total output selected by all other players in the market in round t − 1.
       Table 5 summarizes the estimation results. We provide estimation results for each
treatment separately for ease of exposition. To test for treatment differences, we ran the
same model estimation where all regressors were interacted with a set of treatment dummies.
The results from that joint estimation can be found in the Appendix.

        We start by examining the restricted version of our model, in which we do not consider
time trends (that is, β4 = β5 = 0). The coefficients on Chat in all three regressions confirm
the analysis of Table 3: communication leads to lower average quantity. We do detect
interesting differences in the three treatments with respect to how individuals reacted to
aggregate quantities in the previous round. In the n = 2 case, we observe a positive and
highly significant coefficient on Q−i,t−1 , and a non-significant interaction of Q−i,t−1 with
Chat. Players therefore respond to higher quantity by their rival in the previous round with
higher quantity in the present round, suggesting a collusive relationship over time between
the two players, and one which does not require communication in order to be effective. In
the n = 4 case, that positive relationship is only present in the presence of communication,
suggesting that collusion is perhaps harder to achieve in larger groups. Finally, in the n = 6
case, we do not observe any relationship between qi,t and Q−i,t−1 either in the presence or
14

absence of communication.
        We next consider the unrestricted model in which we allow for the presence of time
trends, both in the intercept and in its interaction with the Chat dummy, which capture the
extent to which average quantity is allowed to vary over the course of the experiment in the
two communication regimes. In the n = 2 and n = 4, the coefficients on these time trends
are not statistically significant. This suggests that the relationship between qi,t and Q−i,t−1
(if any) is relatively stable over time in both group sizes. The same is not the case for the
n = 6: introducing the time trends increases the absolute size of the coefficient on Chat,
which is now larger. We also observe a negative and significant time trend and a positive
significant interaction with Chat; both are small in magnitude and seem to be dominated
by the intercept effect. That is, the ability to communicate leads to a very large drop in
output, which seems to increase slightly over time. This is congruent with the possibility
that subjects were approaching the collusive output from below. In any event, players react
to past changes in the aggregate output produced by other players in their market in the
past round in a manner consistent with collusion.
        As a caveat, we note that, due to the experimental design, we did not collect subjects’
contemporaneous beliefs about the output choices made by the other firms in a given period.
It is quite reasonable to expect that these beliefs would be an important determinant of
quantity choices. If those beliefs are positively correlated with past choices, then we would
expect the estimated coefficients on Q−i,t−1 to be biased upwards because of omitted variable
bias. In other words, our econometric results might be over-estimating the effect past choices
by other firms have on current output choices.
        Figure 3 illustrates the analysis presented so far. It shows the estimated per round
average quantity for the six treatments. The left panel concerns the No-Chat treatments,
while the right panel concerns the Chat conditions. Conditional on the communication
regime, we observe clear mean differences across treatments, which remain constant over
time. It is quite clear from the estimated 95% confidence intervals that there is much more
variability in per-round output in the No-Chat treatments than in the Chat treatments.

5.3    The gain from communication
So far, what we found suggests (more or less) Nash-equilibrium play without communication
and near-perfect symmetric collusion with chat. This was consistent with Hypotheses 1 (i)
and 2 but not supporting Hypothesis 1 (ii).
        We now put these findings together in terms of the gain from talking formalized in
Hypothesis 3. Table 6 displays the estimation results of a simple treatment-effects interaction
15

                                                     NoChat                                                                                      Chat
   10 15 20 25 30 35 40 45

                                                                                              10 15 20 25 30 35 40 45
      Average Quantity

                                                                                                 Average Quantity
               5

                                                                                                          5
               0

                                                                                                          0
                             1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24                             1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
                                                     Period                                                                                     Period

                                               N=2            N=4          N=6                                                            N=2           N=4           N=6

Figure 3: Figure 3: Time series of average prices. Note: solid lines denote estimated per
round means, and dashed lines indicate 95% confidence intervals. Black, gray and blue
correspond to n = 2, 4, 6, respectively.

                                          P
model with the per round industry profit i πi,t as the dependent variable. We estimate the
model using Random Effects GLS with robust standard errors clustered at the market level.
The difference in profits resulting from communication, ∆Π, is reflected by the interaction
terms n × Chat. This difference is higher in n = 4 than in n = 2 (χ2 (1) = 8.41, p = 0.004).
It is also higher in n = 6 than in n = 4 (χ2 (1) = 19.04, p < 0.0001). This is support for
Hypothesis 3 which suggests that ∆Π increases in n.14

Observation 3: The gain from communication, ∆Π, (i) increases in n, and it does so (ii)
at a decreasing rate.

       It is informative to compare the results from this experiment to those of Fonseca and
Normann (2012), who study the effect of communication in Bertrand markets in markets
with two, four, six and eight firms. Unlike our paper, Fonseca and Normann (2012) find
an inverted-u relationship between the number of firms and the absolute gains from com-
munication. The difference in results is a function of what firms do when communication is
available. In the Bertrand environment without communication, average prices are collusive
for n = 2 but they are very close to Nash for n = 4 and n = 6. We observe the same pattern
in the Cournot environment. Communication was very effective in the Cournot markets ir-
respective of the number of firms, as the average posted quantity was very close to the joint
  14
    Observation 3, which can be seen as our main result, is already reflected in average euro payments. When
we calculate the difference in average profit per period per participant, explicit communication increased
profit per period per participant by 0.04 euro in the n = 2 markets, by 0.21 euro in the n = 4 markets,
and by 0.29 euro in the n = 6 markets, highlighting our conclusion that the gain from explicit collusion is
increasing in n, but at a decreasing rate.
16

                                                             P
                          DV:                                    πi,t
                          n = 2 × Chat                      167.58∗∗∗
                                                              (48.95)
                          n=4                               -716.28∗∗∗
                                                             (149.21)
                          n = 4 × Chat                      840.61∗∗∗
                                                             (144.52)
                          n=6                              -1275.92∗∗∗
                                                              (90.75)
                          n = 6 × Chat                     1307.74∗∗∗
                                                             (112.15)
                          Constant                         2261.32∗∗∗
                                                              (48.57)
                          R2                                  0.51
                          N                                   1,152
                          ∗∗∗ ∗∗ ∗
                            , , : p < 0.01, p < 0.05, p < 0.10.
                          Robust SEs clustered at the market level in parentheses.

Table 6: Random effects GLS estimates of firm-level and market-level profits as a function
of number of firms

profit maximising level. In contrast, the Bertrand environment led cooperation to break
down under communication as firm numbers increased. There are two possible reasons for
this discrepancy in results. While the monopoly price in the Bertrand game studied by Fon-
seca and Normann (2012) was a natural focal point, any deviation by at least one firm led to
all other firms earning zero profit for the period. In contrast, the quadratic profit function
in the Cournot game means that firms have a harder task to find the joint-profit maximum
quantity from a cognitive perspective, but deviations from the collusive equilibrium are less
punitive for the cheated firms.

5.4       Text-mining analysis of the communication data
In this section we analyze the language used in the Chat variants in more detail. In particular,
we wish to understand the extent to which the number of firms affects the language used,
and what kind of language is useful to support collusion. To this effect, we employ text
mining methods (Moellers, Normann and Snyder, 2017).15 Text-mining methods extract
keywords from a body of text, referred to as a corpus. We will compare the most frequently
used keywords for two corpora in order to find out how the corpora (the chats) differ. To
be more precise, we will use Huerta’s (2008) relative rank difference which tells us which
 15
      For alternative methods of text analysis, see Kimborough et al. (2008) or Houser and Xiao (2011).
17

keywords are comparatively more frequently used in corpus c relative to c0 . Formally, we
measure the keyness of word w in corpus c relative to c0 by generating ranks rc (w) for all
words w in corpus c according to frequency (and in descending order). The difference in the
rank of w in corpus c relative to corpus c0 is defined as

                                       0    |rc0 (w) − rc (w)|
                                   rdcc =                      .
                                                   rc (w)

As Huerta (2008, p. 967) points out, the rd score “denotes some sort of percent change in
rank. This also means that this function is less sensitive to small changes in frequency in the
case of frequent words and to small changes in rank in case of infrequent words.” In other
words, we are not concerned about cardinal measures, but ordinal measures, making the
analysis distribution-free. This measure means that a unit change on the top of the ranking,
say from first to second, will have a higher rd score than a change from 50th to 51st. This is
intuitive because changes at the top of the ranking are more important than changes at the
bottom of the ranking, since the former apply to very frequently-used words, as opposed to
the latter. Had we wanted to give changes a more equal weighting throughout the ranking,
we would have raised the denominator to the power of larger than one. Also, a large rd score
implies a large leap in the rankings from one corpus to the other, meaning that it is almost
never used in one context to being very frequently used in the other.
        In our analysis, we always compare the difference in the rank of w in corpus c relative
to corpus c0 and corpus c0 relative to corpus c to get a complete picture. We restrict ourselves
to keywords that are among the top 50 most common in corpus c, avoiding keywords with a
         0
high rdcc that are nevertheless rarely used. We omit conjunctions, prepositions, and articles
                                         0
and we only report keywords with rdcc > 1.
        Table 7 reveals some interesting insights into the differences in chat when it comes
to the number of firms. It is instructive to look at the words that have the highest rank
differential in the pairwise comparisons. In duopolies, ‘25’ is discussed relatively more often
than in the other market structures; ‘12’ is relatively more frequent in four-firm markets;
and ‘8’ is relatively more frequently used in six-firm markets than elsewhere. This is hardly
surprising: these numbers are the joint-profit maximizing outputs. It illustrates that subjects
identified what the profit-maximizing output was, and attempted to coordinate on that value.
The relatively high frequency of other close values could be an indication of learning or trial
and error — recall that subjects did not have a payoff table. Furthermore, markets with
n = 4 and n = 6 cannot produce the joint-profit maximizing aggregate output of 49 or 50
symmetrically. Hence, they also talk about targets other than ‘12’ or ‘8’.
two vs. four                      two vs. six                       four vs. six
          n=2            n=4                  n=2          n=6                 n=4            n=6
       word  rd      word    rd            word   rd   word   rd             word   rd word      rd

         25  261.0       12      716.0        25    581.0     8     606.0      12   70.0      8    151.0
         26   50.5       11       78.7        26    51.9      9     357.5      13   20.6      7    50.6
       both 38.1         13       64.2        24     36.5     7      26.1      11   10.2      9     15.5
         24   35.6       10       27.8        23    31.3     all     23.0      :)   4.0      B      4.9
         27   27.4        9      20.7         27    28.1     B       22.1      20    3.9      5     3.2
         23   21.7    everyone   11.0       both    15.4     10      22.0   oneself 3.1     one     3.0
         50   16.5       15       4.9          I     5.0     A      16.9s   already 2.3       6     2.5
         let   7.8      each       2.7     times      4.0    D       4.6      each  1.3     stay    2.2
           I   5.0     works       2.0        50     3.8      5      3.1     times   1.1     do     2.1
       think   3.8      gives     1.6        you     2.4    one      2.5                    but     1.2
       have    2.9        :)       1.5       let     2.4     do       1.8                    A      1.1
        stay  2.3       does      1.3         20     2.0    only     1.7
        you    2.1      how       1.1      think     1.8     to      1.4
          or  1.9                             ;)     1.7    still    1.1
       times 1.4                            still    1.7
        but    1.1                         better    1.5
         for   1.1                            or     1.5
               1.1                          than     1.2
               1.1                          have     1.2

Table 7: Text-mining analysis. We report words with absolute rank rc ≤ 50 and relative rank differential rd ≥ 1.

                                                                                                                   18
19

        A conspicuous finding between duopolies on the one hand and the four- and six-firm
treatments on the other is the relatively more frequent use of ‘both’ for n = 2, and the
relatively more frequent use of ‘everyone’, ‘one’, or ‘all’ for n = 4 and n = 6.16 These words
were presumably used in the context of invoking a collective decision, or attempting to invoke
a group identity. It is interesting to note that ‘I’ and ‘you’ were used more often in duopolies
than in the other two treatments. This could signify that subjects may have attempted to
coordinate on strategies that involved asymmetric output choices. In contrast, subjects the
six-firm markets used capital letters (individual subjects were identified by capital letters on
the screen) more. While duopolists may not have needed to use letters (hence their frequent
usage of ‘you’), the use of letters in a large group indicates a particular individual was singled
out by participants. One conjectures this was done as a reprimand for bad behavior in a
previous round, or perhaps less likely, as a compliment for abiding to an agreement.
        Another interesting target for language analysis is to compare groups that successfully
colluded to groups that did not. A problem is that virtually all groups can be considered
collusive. For example, 21 of 24 groups have an average quantity of plus ten percent on top
of the joint-profit maximizing output. And even the remaining three groups (all of which
have n = 6) have average outputs closer to the joint profit maximum than to the static
Nash equilibrium. The text-mining analysis still leads to some insights. We produce the
comprehensive analysis in Table 9 in the Appendix and mention some conspicuous findings
here. The successful groups (according to the above criterion), unsurprisingly, mention more
frequently the collusive quantity targets “13”, suitable for n = 4, and “25”, suitable for
n = 2. The three “unsuccessful” groups used relatively frequently non-suitable (for n = 6)
quantity targets like “7” or “10”, and “please”, indicating disagreement or uncertainty about
choices.
        In Table 10 in the Appendix, we also provide a ranking of the words used most
frequently in absolute terms. The table shows that, unsurprisingly, some words are frequently
used throughout (“I”, “ok”), but the quantities of the symmetric collusive outputs also
appear at the top of the list. Since the collusive outputs differ across treatments, they have
a particularly high rd score. For example, the quantity “12” is the most frequently used
term with n = 4 but is never used with n = 2. These absolute frequencies imply a rather
large relative rank difference of rdn=2n=4 = 716.0. Intuitively, also more moderate differences
in ranks can imply a substantial rd score when they score high in absolute frequency. “I”
is the most frequent word with n = 2 and it is the sixth most frequent word with n = 4,
suggesting rdn=4
              n=2 = 5.0, that is, the relative rank difference is five.
 16
      The “one” in the list of words refers to the German “einer,” as in “one of us”.
20

6    Discussion
The main research question of this paper is to quantify firms’ additional profit from talking
explicitly for Cournot oligopolies. We observe an increasing and concave relationship between
the number of firms and this gain. In other words, markets with more firms find it more
profitable to talk than markets with fewer firms.
        Our finding is, on the one hand, consistent with the meta study of Balliet (2010). On
the other hand, Fonseca and Normann (2012) found an inverse-u shaped relationship the
authors should add between the number of firms and the incentive to collude. This shows
that strategic substitutes vs. complements may matter (see also Mermer, Müller and Suetens,
2016, for duopolies without talk) regarding the incentive to talk, as may asymmetries between
firms (see Harrington et al. 2016). In any event, it cannot be taken for granted that the
conventional wisdom—fewer firms find it easier to collude—reflects the gain from explicitly
talking and therefore the frequency of cartels.
        We furthermore find evidence confirming the role of communication as a catalyst
to cooperation in repeated market games. Communication leads to a reduction in average
quantities, and lower dispersion, suggesting that it facilitates coordination. The dynamics of
output choices shows there are differences in the way communication helps collusion as the
number of firms increases. In duopolies, the effect of communication primarily materializes
through a level effect. We find dynamic output adjustments where firms positively respond
to the quantity posted by the other firm in the previous period. Those adjustments are equal
in magnitude in the Chat and No-Chat conditions. In contrast, the data from the four-firm
markets shows dynamic quantity adjustments in the Chat conditions only. Coordinating on
the profit maximizing output is a more difficult proposition when done by four firms, and
communication appears to help. A still different pattern emerges in the six-firm markets.
There, we find no dynamic output adjustments in response to other firms’ behavior. Instead,
we find a simple time trend effect, suggesting that communication only allowed subjects to
recognize the advantages of lower quantities, while not allowing them to adjust outputs
optimally — which is consistent with the observation that aggregate output in the six-firm
markets with communication were substantially below the joint-profit maximum.
21

References
[1] Aubert, Cecile, Patrick Rey, and William E. Kovacic. 2007. The Impact of Leniency and
   Whistleblowing Programs on Cartels. International Journal of Industrial Organization
   24(6): 1241-1266.

[2] Balliet, Daniel. 2010. Communication and Cooperation in Social Dilemmas: A Meta-
   Analytic Review. Journal of Conflict Resolution 54(1): 39-57.

[3] Belleflamme, Paul and Martin Peitz. 2015. Industrial Organization, Markets and Strate-
   gies. Cambridge University Press, Cambridge.

[4] Binger, Brian R., Elisabeth Hoffman, Gary D. Libecap. 1990. An experimetric study of
   the Cournot theory of firm behavior. Working paper, University of Arizona.

[5] Boone, Jan, Wieland Müller, and Sigrid Suetens. 2014. Naked Exclusion in the Lab: The
   Case of Sequential Contracting. Journal of Industrial Economics 62(1): 137-166 .

[6] Brosig, Jeannette, Axel Ockenfels, and Joachim Weimann. 2003. The Effect of Commu-
   nication Media on Cooperation. German Economic Review 4: 217-241.

[7] Cason, Timothy N., and Douglas D. Davis. 1995. Price Communication in a Multi-Market
   Context. Review of Industrial Organization 10: 769-787.

[8] Davis, Douglas D., and Charles A.Holt. 1994. Market Power in Laboratory Markets with
   Posted Prices. The Rand Journal of Economics 25: 467-487.

[9] Davis, Douglas D., and Charles A. Holt. 1998. Conspiracies and Secret Price Discounts.
   Economic Journal 108: 1-21.

[10] Deutsch, Morton. 1958. Trust and suspicion. Journal of Conflict Resolution 2(4): 265-
   279.

[11] Dolbear, Trenery F., Lester B. Lave, G. Bowman, G., A. Lieberman, A., Edward C.
   Prescott, F. Rueter, and Roger Sherman. 1968. Collusion in Oligopoly: An Experiment on
   the Effect of Numbers and Information. Quarterly Journal of Economics 82(2): 240-259.

[12] Fischbacher, Urs. 2007. z-Tree - Zurich toolbox for Readymade Economic Experiments.
   Experimental Economics 10(2): 171-178.

[13] Fonseca, Miguel A. and Hans-Theo Normann. 2012. Explicit vs. Tacit Collusion: The
   Impact of Communication in Oligopoly Experiments. European Economic Review 56:
   1759-1772.
22

[14] Fouraker, Lawrence and Sidney Siegel. 1963. Bargaining Behavior, McGraw-Hill, New
   York.

[15] Friedman, James W. 1967. An Experimental Study of Cooperative Duopoly. Economet-
   rica 35: 379-397.

[16] Gomez-Martinez, Francisco, Sander Onderstal, and Joep Sonnemans. 2016. Firm-
   Specific Information and Explicit Collusion in Experimental Oligopolies. European Eco-
   nomic Review 82: 132-141.

[17] Greiner, Ben. 2015. Subject Pool Recruitment Procedures: Organizing Experiments
   with ORSEE. Journal of the Economic Science Association 1(1): 114-125.

[18] Grout, Paul A. de, and Silvia Sonderegger. 2005. Predicting cartels. Office of Fair
   Trading Economic Discussion Paper.

[19] Harrington, Joseph E. 2008. Detecting Cartels. In Handbook in Antitrust Economics,
   ed. Paolo Buccirossi, 213-258. MIT Press, Cambridge (Mass.).

[20] Harrington Jr, Joseph E, Roberto Hernan Gonzalez, and Praveen Kujal. 2016. The
   Relative Efficacy of Price Announcements and Express Communication for Collusion:
   Experimental Findings. Journal of Economic Behavior and Organization 128: 251–264.

[21] Hay, George A. and Daniel Kelley. 1974. Empirical Survey of Price Fixing Conspiracies.
   Journal of Law and Economics 17: 13-38.

[22] Holt, Charles A. 1985. An Experimental Test of the Consistent-Conjectures Hypothesis.
   American Economic Review 75(3): 314-325.

[23] Holt, Charles A. and Douglas D. Davis. 1990. The Effects of Non-Binding Price An-
   nouncements in Posted-Offer Markets. Economics Letters 34(4): 307-310.

[24] Houser, Dan, and Erte Xiao. 2011. Classification of Natural Language Messages Using
   A Coordination Game. Experimental Economics 14: 1–14.

[25] Huck, Steffen, Hans-Theo Normann, and Joerg Oechssler. 2004. Two are Few and Four
   are Many – on Number Effects in Cournot Oligopoly. Journal of Economic Behavior and
   Organization 53: 435-446.

[26] Huerta, Juan M. 2008. Relative Rank Statistics for Dialog Analysis. Proceedings of the
   2008 Conference on Empirical Methods in Natural Language Processing, Association for
   Computational Linguistics: 965-972.
23

[27] Isaac, R. Mark, Valerie Ramey, and Arlington W. Williams. 1984. The Effects of Market
   Organization on Conspiracies in Restraint of Trade. Journal of Economic Behavior and
   Organization 5: 191-222.

[28] Isaac, R. Mark and James M. Walker. 1988. Communication and Free-Riding Behavior:
   The Voluntary Contribution Mechanism. Economic Inquiry 26(4): 585-608.

[29] Ivaldi, Marc, Bruno Jullien, Patrick Rey, Paul Seabright, and Jean Tirole. 2003. The
   Economics of Tacit Collusion. Report for DG Competition, European Commission.

[30] Kimbrough, Erik, O., Vernon Smith, and Bart Wilson. 2008. Historical Property Rights,
   Sociality, and the Emergence of Impersonal Exchange in Long-Distance Trade. American
   Economic Review 98(3): 1009-1039.

[31] Landeo, Claudia M., and Kathryn E. Spier. 2009. Naked Exclusion: An Experimental
   Study of Contracts With Externalities. American Economic Review 99: 1850-1877.

[32] Levenstein, Margaret C. and Valerie Y. Suslow. 2006. What Determines Cartel Success?
   Journal of Economic Literature 44: 43-95.

[33] Li, Yan and Bruce Lyons. 2012. Market Structure, Regulation and the Speed of Mobile
   Network Penetration. International Journal of Industrial Organization 30(6): 697-707.

[34] Martin, Stephen. 2001. Advanced Industrial Economics, Second edition, Blackwell Pub-
   lishers.

[35] Mermer, Ayse Gül, Wieland Mueller, and Sigrid Suetens. 2016. Cooperation in Indef-
   initely Repeated Games of Strategic Complements and Substitutes. Working Paper No
   1603, Vienna University.

[36] Motta, Massimo. 2004. Competition Policy: Theory and Practice, Cambridge University
   Press, Cambridge.

[37] Normann, Hans-Theo, Jürgen Rösch, and Luis-Manuel Schultz. 2015. Do Buyer Groups
   Facilitate Collusion? Journal of Economic Behavior and Organization 109: 72-84.

[38] Posner, Richard A. 1970. A Statistical Study of Antitrust Enforcement. Journal of Law
   and Economics 13(2): 365-419.

[39] Scherer, Frederik M. 1980. Industrial Market Structure and Economic Performance.
   Rand McNally College Pub. Co., (Chicago).
24

[40] Tirole, Jean. 1989. The Theory of Industrial Organization. MIT Press, Cambridge
   (Mass.).

[41] Waichman, Israel, Till Requate, Chng Kean and Siang. 2014. Pre-play Communica-
   tion in Cournot Competition: An Experiment with Students and Managers. Journal of
   Economic Psychology 42: 116

[42] Whinston, Michael D. 2008. Lectures on Antitrust Economics. MIT Press, Cambridge
   (Mass.).
25

Appendix 1: General Model
Underlying our design is a homogenous-good Cournot oligopoly with n firms as players.
Firms choose quantities qi ∈ [0, ∞), i = 1, ...n. Let Q = ni qi . Inverse demand is linear such
                                                         P

that p = max{a − Q, 0}, where p is the market price. Firms produce at constant marginal
costs of c. Profits are denoted by Πi = (p − c)qi .
       The two benchmarks we use are the static Nash equilibrium and the symmetric joint-
profit maximum. In the static Nash equilibrium, each firm produces qi = (a − c)/(n + 1)
and earns a profit of                             2
                                            a−c
                                                      .                                     (3)
                                            n+1
For n = 1, we obtain the monopoly output, qi = (a − c)/2. The symmetric joint-profit
maximum has therefore each of the n firms producing qi = (a − c)/2n, yielding a profit for
each firm equal to
                                             (a − c)2
                                                      .                                     (4)
                                                4n
        We now show how the joint-profit maximum can be sustained as a subgame perfect
Nash equilibrium (SGPNE) in a repeated game. Consider an infinitely repeated version of
the above stage game. Firms discount future profits by a factor δ ∈ (0, 1). Suppose that
firms aim at maintaining the symmetric joint payoff maximum as an SGPNE with a simple
Nash trigger strategy. If firm i deviates, its best response is to produce (a − c) (n + 1) /4n,
yielding a defection profit of ((a − c) (n + 1) /4n)2 . For the symmetric joint payoff maximum
to be an SGPNE, the stream of discounted collusive profits has to be at least as high as the
profit from a one-time deviation followed by a grim (Nash) punishment path in the future,
that is:
                        (a − c)2      (a − c)2 (n + 1)2        δ(a − c)2
                                   ≥                     +                                  (5)
                       (1 − δ)4n            16n2           (1 − δ)(n + 1)2
We can solve this inequality for δ. The discount factor has to be at least

                                           (n + 1)2
                                   δ≥                 ≡δ                                   (6)
                                        (n + 1)2 + 4n

for cooperation to be a subgame perfect equilibrium.
       We find ∂δ/∂n > 0, so the minimum discount factor increases in n. The incentive
constraint (6) is a condition that necessarily has to be met in repeated games, regardless
of whether players communicate explicitly. Furthermore, conditions like (6) are often inter-
preted as an indicator of how “difficult” collusion is. Thus we have theoretical support for
Hypothesis 1 in the main text. We note that Hypothesis 1 holds for other collusive equi-
You can also read